[Bug]: Move enum explanations and health checks from cuda_core to cuda_bindings

[rwgk]

Related: nvbug 5932944

Situation

Currently (main @ eabaf4b) we have these enum explanations and related health checks :

• https://github.com/NVIDIA/cuda-python/blob/main/cuda_core/cuda/core/_utils/driver_cu_result_explanations.py

• https://github.com/NVIDIA/cuda-python/blob/main/cuda_core/cuda/core/_utils/runtime_cuda_error_explanations.py

• cuda-python/cuda_core/tests/test_cuda_utils.py

  Lines 11 to 40 in eabaf4b

  	def test_driver_cu_result_explanations_health():
  	expl_dict = cuda_utils.DRIVER_CU_RESULT_EXPLANATIONS
  	# Ensure all CUresult enums are in expl_dict
  	known_codes = set()
  	for error in driver.CUresult:
  	code = int(error)
  	assert code in expl_dict
  	known_codes.add(code)
  	if cuda_utils.get_binding_version() >= (13, 0):
  	# Ensure expl_dict has no codes not known as a CUresult enum
  	extra_expl = sorted(set(expl_dict.keys()) - known_codes)
  	assert not extra_expl
  	def test_runtime_cuda_error_explanations_health():
  	expl_dict = cuda_utils.RUNTIME_CUDA_ERROR_EXPLANATIONS
  	# Ensure all cudaError_t enums are in expl_dict
  	known_codes = set()
  	for error in runtime.cudaError_t:
  	code = int(error)
  	assert code in expl_dict
  	known_codes.add(code)
  	if cuda_utils.get_binding_version() >= (13, 0):
  	# Ensure expl_dict has no codes not known as a cudaError_t enum
  	extra_expl = sorted(set(expl_dict.keys()) - known_codes)
  	assert not extra_expl

Problem

The tests succeed only for the matching cuda-bindings release.

If cuda_core is tests are run against an older or newer version of cuda-bindings, they will fail (that's nvbug 5932944).

Solution

The explanations dicts are fundamentally directly tied to the bindings, therefore they should be moved to cuda_bindings (maybe a new cuda_bindings/cuda/bindings/_utils/ directory). Accordingly, the tests should be moved, too.

Backward compatibility

When the explanations dicts are moved, the cuda_core code using them needs to be made tolerant to not finding them in existing releases of cuda-bindings. I.e. the explanations will be used if available in cuda_bindings, or otherwise simply not be added when generating error messages at runtime. This is considered to be a minor loss that will become a non-issue over time as old cuda-bindings releases fall out of use.

[leofang]

Make sure that

• if the explanation dict exists in cuda-bindings, use it
• otherwise, use the vendored copy in cuda-core

We still support older versions of cuda-bindings, so need to be mindful.

[rwgk]

otherwise, use the vendored copy in cuda-core

I was thinking it's better to not vendor; i.e. vendoring is more trouble than it's worth. The corresponding part of the Backward compatibility section: "I.e. the explanations will be used if available in cuda_bindings, or otherwise simply not be added when generating error messages at runtime. This is considered to be a minor loss that will become a non-issue over time as old cuda-bindings releases fall out of use."

@leofang do you think I should reconsider?